CAPÍTULO V. EFECTO DE LOS MEDIOS AUDIOVISUALES 113
V. 11. Percepción selectiva: cada uno ve lo que quiere ver 144
2.1. Acquire data to develop input layers for WXFIRE.
2.1.1. Acquire Digital Elevation Model (DEM) layer to create elevation, aspect, slope, and topographic shading layers.
2.1.2. Acquire STATSGO soils coverage and associated tabular data. 2.1.3. Acquire NLCD layer to create Ecophysiological Site layer. 2.1.4. Acquire DAYMET weather database.
2.1.5. Acquire Landsat imagery for leaf-on reflectance date to create Leaf Area Index (LAI) layer.
2.2. Create terrain-related layers.
2.2.1. Create Slope layer using Arc/Info SLOPE command with PERCENTRISE as units of slope. 2.2.2. Create Aspect layer using Arc/Info ASPECT command.
2.2.3. Create Topographic Shading layer using Arc/Info HILLSHADE command. (Azimuth and altitude data were developed using NOAA Solar Position calculator, assuming summer solstice as the date and using center coordinates for each zone.)
2.3. Create Soil Texture layers (percent sand, percent silt, percent clay).
2.3.1. Using STATSGO database, compute four soil textures (percent sand, percent silt, percent clay, and coarse fragment).
2.3.2. Weight each soil texture by the layers’ depths and spatial extent for each of soil sequences within STATSGO polygons.
2.3.3. Remove coarse fragment proportion from the composition of soil textures and rescale sand, silt, and clay components to comprise 100 percent of soil texture estimates.
2.3.4. Calculate average slope in STATSGO database from high and low values for each STATSGO poly- gon and associated sequences and classify average slope into 4 classes:
(1) ≤ 4 percent; (2) >4 percent and ≤8 percent; (3) > 8 percent and ≤ 15 percent; and (4) >15 percent.
2.3.5. Calculate average soil texture using data from step 2.3.3 for each slope class within each STATSGO polygon.
2.3.6. Classify Slope layer (from 2.2.1) into same 4 slope classes.
2.3.7. Partition STATSGO polygon coverage by Classified Slope layer and link this spatial layer with the
STATSGO variables of soil texture by polygon and slope class (from 2.3.5). 2.4. Create Soil Depth layer.
2.4.1. Extract data on maximum depth per soil sequence from the STATSGO database.
2.4.2. Weight maximum depth per soil sequence by areal extent of sequences to calculate maximum soil depth per polygon.
2.4.3. Calculate Topographic Soil Index (TSI) for each pixel using the following relationship:
TSI= ln aB tan
where a is upslope area (m2) draining past a certain point per unit width of slope calculated using
Arc/Info FLOWACCUMULATION and FLOWDIRECTION commands and B is local surface
slope angle (degrees) calculated using Arc/Info SLOPE command with DEGREE as units of slope. 2.4.4. Integrate STATSGO Maximum Depth layer and TSI to calculate soil depth value for each pixel
using scalars to adjust for skewed TSI distributions in the equation: Soil Depth = {M1, M2} *TSI.
where M1 is scalar used if pixel’s TSI is ≤ mean across a mapping zone, and M2 is used if TSI
value is > mapping zone’s mean.
Appendix 2-A — (Continued)
Calculate M1 and M2 by the formulas:
M
LNmo LNme M
1=0.5 (Ave. Max. Depth∗ + ) and 2= MaLNxx. Depth
max
where ave. max. depth is mean value of the STATSGO Maximum Depth layer across each zone, and LNmoand LNme are the mode and mean of the natural log of TSI for each STATSGO polygon
calculated using Arc/Info’s ZONALMAJORITY and ZONALMEAN commands, respectively.
2.4.5. For Zone 19: increase data resolution using slope data from STATSGO database and Classified
Slope layer.
2.4.5.1. Use slope classes calculated from STATSGO database in step 2.3.1.
2.4.5.2. Calculate average maximum depth for each slope class within each STATSGO polygon using data from step 2.3.2.
2.4.5.3. Link STATSGO polygon coverage partitioned by Classified Slope layer from step 2.3.7
with STATSGO average maximum depth by polygon and slope class data calculated in 2.4.5.2.
2.5. Create LAI layer.
2.5.1. Calculate corrected Normalized Difference Vegetation Index (NDVI) using LANDSAT leaf-on
reflectance imagery and the equation:
NDVIc=NIR REDNIR RED+− ∗ −1 MINMIR MIR− min
maax− min MIR
where NIR is near infrared (band 4), RED is infrared (band 3), and MIR is mid-infrared (band 5); MIRmin is minimum value in mid-infrared band in an open canopy; and MIRmax is maximum value in
the mid-infrared band in a closed canopy.
2.5.2. Convert NDVIc layer to LAI using the equation:
LAI= −NDVIc − 1 0 7 0 7 n( . . 2.6. Create Weather layer.
2.6.1. Using any one of the DAYMET layers (for example, daily temperature), clip DAYMET layer to zonal boundary using Arc/Info GRIDCLIP command.
2.6.2. Use clipped DAYMET layer to obtain center coordinates for each 1-km pixel. 2.7. Create Ecophysiological Site layer.
2.7.1. For Zone 16, partition landscape by 4 elevational breaks using DEM: Site 1 – 0 to 4,000 ft mean sea level (MSL); Site 2 – 4,000 to 6,000 ft MSL; Site 3 – 6,000 to 9,000 ft MSL; and Site 4 – 9,000+ ft MSL.
2.7.2. For Zone 19, reassign 21 broad CTs from NLCD to 4 general plant functional types and one non- vegetated class: water/barren. Reassign developed land CTs to plant functional types based on surrounding pixels using FOCALMAJORITY command.
2.8. Classify WXFIRE input layers.
2.8.1. Classify Elevation layer into 100-m ranges.
2.8.2. Classify Slope layer (from 2.2.1) into low (0-10%), moderate (10-30%), and high (>30%) slope
classes.
2.8.3. Classify Aspect layer into SW (165° to 255°), NW (255° to 345°), NE (345° to 75°), and SE (75° to 165°) classes.
2.8.4. Classify Topographic Shading Index layer into 0.25 intervals. 2.8.5. Classify Soil Depth layer into 0.5-m intervals.
2.8.6. Classify LAI layer into 1.0 intervals.
2.9. Create simulation units for running WXFIRE model.
2.9.1. Combine classified input layers (terrain, soil depth, and LAI), and ecophysiological site and weath- er layers such that each unique combination forms one simulation unit using Arc/Info’s COMBINE command.
2.9.2. Associate values from each input layer to each simulation unit.
2.9.3. Create ASCII file for input to WXFIRE model that lists all the simulation units in a mapping zone
with their associated site, terrain, weather-coordinates, soils, and LAI values. 2.10. Run WXFIRE simulations and develop biophysical gradient layers.
2.10.1. Input ASCII file to WXFIRE model.
2.10.2. Link each record in ASCII output file from WXFIRE model to its geo-referenced simulation unit
(from step 2.9).
2.10.3. Create individual biophysical gradient layers for each simulation unit. 3. Mapping potential vegetation type (PVT)
3.1. Prepare data for model building.
3.1.1. Prepare spatially explicit predictor layers (biophysical and topographic gradients). 3.1.1.1. Acquire biophysical and topographic gradients for 3-km buffered zone.
3.1.1.2. Scale all layers to unsigned 8-bit or 16-bit integers and output summary statistics for each layer.
3.1.1.3. Convert layer to unsigned 8-bit or 16-bit integer images.
3.1.1.4. Quality-check all predictor layers.
3.1.1.4.1. Check projections and row / column numbers for consistency. 3.1.1.4.2. Check all images for erroneous numbers or patterns.
3.1.2. Prepare response data (PVT classes).
3.1.2.1. Acquire LFRDB MAT with uniqueID, spatial reference, and PVT assignments for plots within zone boundary.
3.1.2.2. Examine data spatially and non-spatially, looking for outliers or unusual spatial distribu- tions.
3.1.2.3. Evaluate number of available plots by PVT class to see if classes need to be collapsed or dropped.
3.1.2.4. Label each PVT plot as forest or non-forest type using values 1 and 2, respectively. 3.1.3. Perform data extraction.
3.1.3.1. Extract values from each predictor gradient for each X and Y plot coordinate and link to the LFRDB MAT.
3.1.4. Perform data exploratory exercises.
3.1.4.1. View data spatially, looking for unusual spatial patterns or outliers.
3.1.4.2. Import data into a statistical package (in other words, R) and examine data for outliers or unusual features.
3.1.4.2.1. Examine summary statistics of response (box plots, etc.).
3.1.4.2.2. Examine summary statistics of predictors (distributions, scatter plots, correlation matrices, and principal components).
3.2. Generate PVT life form (forest / non-forest) model and map.
3.2.1. Set up input files for the See5 application.
3.2.1.1. Generate an ERDAS Imagine image (dependent variable) of training plots using forest / non-forest values.
3.2.1.2. Use NLCD Mapping Tool and Sampling Tool to generate See5 .names input file. 3.2.1.3. Delete .data and .test files that are output from the NLCD Sampling Tool.
3.2.1.4. Export refined training data set to a comma-delimited file (.data) including the uniqueID, the predictor gradient values (in the same order as listed in the .names file) and dependent
(forest / non-forest) value.
3.2.2. Use See5 to build forest / non-forest model.
3.2.2.1. From See5, open input files (.data and .names).
3.2.2.2. Specify options (such as winnow, boosting, and misclassification cost).
3.2.2.3. Run model with 10-fold cross-validation (for accuracy assessment).
3.2.2.4. Run model without cross-validation (for generating .tree file for prediction).
3.2.3. Apply model across buffered zone.
3.2.3.1. Use NLCD Mapping Tool to generate a Forest / Non-forest map with an associated map of
confidence.
3.3. Extract value from predicted map of forest / non-forest and link to LFRDB MAT. 3.4. Generate 2 mask images of PVT life form (forest / non-forest).
3.4.1. Create a new image by recoding forest / non-forest image to forest – 1; non-forest – 0. 3.4.2. Create a new image by recoding forest / non-forest image to forest – 0; non-forest – 1. 3.5. Generate forest PVT model.
3.5.1. Set up input files for the See5 application.
3.5.1.1. Query data for forest PVTs, where predicted PVT life form is forest (life form = 1). 3.5.1.2. Generate an ERDAS Imagine image (dependent variable) of training plots using forest
PVT values from query.
3.5.1.3. Use NLCD Mapping Tool and Sampling Tool to generate See5 .names file. 3.5.1.4. Delete .data and .test files that are output from the NLCD Sampling Tool. 3.5.1.5. Export a randomly selected 10% of the data set to a comma-delimited *.test file. 3.5.1.6. Export remaining 90% of the data set to a comma-delimited *.data file.
3.5.2. Use See5 to build forest PVT classification tree. 3.5.2.1. From See5, open input files (.data and .names).
3.5.2.2. Specify options (such as winnow, boosting, and misclassification cost). 3.5.2.3. Run model (no cross-validation) to generate .tree file for prediction.
3.5.3. Apply model across buffered zone.
3.5.3.1. Use NLCD Mapping Tool and Classifier Tool to generate a map of forest PVTs with an associated map of confidence using the forest mask to limit prediction extent.
3.6. Generate non-forest (shrub and herbaceous) PVT model.
3.6.1. Set up input files for the See5 application.
3.6.1.1. Query database for non-forest PVTs, where predicted PVT life form is forest (life form = 2). 3.6.1.2. Generate an ERDAS Imagine image (dependent variable) of training plots using non-forest
PVT values from query.
3.6.1.3. Use NLCD Mapping Tool and Sampling Tool to generate See5 .names file. 3.6.1.4. Delete .data and .test files that are output from the NLCD Sampling Tool. 3.6.1.5. Export a randomly selected 10% of the data set to a comma-delimited *.test file. 3.6.1.6. Export remaining 90% of the data set to a comma-delimited *.data file.
3.6.2. Use See5 to build non-forest PVT classification tree. 3.6.2.1. From See5, open input files (.data and .names).
3.6.2.2. Specify options (such as winnow, boosting, and misclassification cost). 3.6.2.3. Run model (no cross-validation) to generate .tree file for prediction.
3.6.3. Apply model across buffered zone.
3.6.3.1. Use NLCD Mapping Tool to generate a map of non-forest PVTs with an associated map of
confidence using the non-forest mask to limit prediction extent. 3.7. Make final maps and assess accuracy.
3.7.1. Combine forest and non-forest maps.
3.7.2. Combine forest and non-forest error matrices.
3.7.3. Calculate accuracy measures (for example, percent correctly classified, user and producer
accuracy, and Kappa statistic). 4. Mapping existing vegetation
4.1. Conduct spatial QA/QC of field plot data
4.1.1. Conduct QA/QC for non-Forest Inventory Analysis (FIA) data point identification.
4.1.1.1. Convert map attribute coordinate data to point attribute (vector) data. 4.1.1.2. Intersect vector coverage with NDVI Change layer
4.1.1.3. Populate table with NDVI difference values. Large differences in NDVI values are likely to represent plots without recent major vegetation change. (such as ± 2 std dev. from mean NDVI value for table).
4.1.1.4. Identify plots with a “distance to road” of > 30m.
4.1.1.5. If NLCD data for 2001 is available, compare CTs to NLCD classes to check for matches. If NLCD 2001 data is not available, try NLCD 1992 data (provided in LFRDB).
4.1.1.6. Flag values in MAT that require attention based on analyses performed in 4.1. 4.1.2. Identify questionable plots.
4.1.2.1. Overlay points onto imagery stratified by CTs.
4.1.2.1.1. Identify and flag points on roads or other similar types of locations (such as urban
or agriculture) that should not be used for training.
4.1.2.1.2. Identify and flag those points that indicate change has occurred since the field
data were obtained.
4.1.2.1.3. Identify plots with forest CTs located in relatively intact non-forest locations (and vice versa).
4.1.2.1.4. Identify plots typed as conifer located in relatively intact deciduous forest (and vice versa).
4.1.2.2. Flag questionable plots in MAT and omit from future analyses.
4.1.3. Develop a modified MAT storing only field plots that pass the QA/QC process in 4.1.2.
4.1.4. Conduct QA/QC for FIA data (same general process as in 4.1.1 but requires FIA analyst). 4.1.5. Isolate 2% of the sample points to be used for accuracy assessment using the 3x3 km, 2% block
design. 4.2. Preprocess imagery.
4.2.1. Ensure that Landsat imagery used for LANDFIRE mapping is processed to the following specifi- cations:
4.2.1.1. For each path/row, acquire and process 3 seasonally separate dates (spring, summer, and autumn) of Landsat scenes
4.2.1.2. Conduct geometric rectification to terrain precision correction level, resulting in less than
± 15m root mean square error (RMSE) spatial accuracy.
4.2.1.3. Conduct radiometric normalization to calibrate radiance values to at-satellite reflectance
values.
4.2.1.4. Calculate NDVI and tasseled cap transformation values for each of the three dates of the data.
4.2.1.5. Develop preliminary maps of forest, shrub, and herbaceous CTs using methods listed in 3 (potential vegetation mapping) Provide the preliminary maps to the PVT mappers and vegetation modelers for internal use.
4.2.2. Ensure that the PVT map and PVT probability layers are stored in data library
4.2.3. Conduct visual quality check on the PVT layers to ensure no obvious seam lines, dropped pixels, or other quality problems exist.
4.2.4. Assemble imagery, topographic data, biophysical gradient layers, PVT probability layers, and riparian-wetland mask (if available).
4.3. Map life form-specific CT
4.3.1. Extract digital values from the spatial layers (4.2.4) using field plots that have passed the visual
QA/QC inspection process (4.1.3 and 4.1.4).
4.3.2. Determine if a “hierarchical approach” (mapping by high-level stratifications) is needed: if there are strong environmental differences between life form-specific CT classes, consider taking the
hierarchical approach. For example, stratify desert shrub CTs from upland and riparian shrub CTs. If the hierarchical approach is needed, go to 4.3.2.1; otherwise, go to 4.3.3.
4.3.2.1. Recode field plot data to high-level CT groups and run decision tree model for high-level
CT groups.
4.3.2.2. Model CTs with decision tree model under each of the high-level CT groups.
4.3.2.3. Calculate overall cross-validation accuracy by weighting and summarizing all CT groups 4.3.2.4. If weighted cross-validation is satisfactory, merge all CT groups into one CT map by major
life form.
4.3.2.5. If weighted cross-validation is not satisfactory, consider rearranging high level groups or abandoning the approach.
4.3.3. Run decision tree model separately for forest, shrub, and herbaceous life forms.
4.3.4. Generate life form-specific cross-validation error matrices.
4.3.5. Generate life form-specific CT layers by applying decision tree models (create separate tree, shrub,
and herbaceous layers).
4.3.6. Check for any visual and information content problems by examining CT maps and interpreting error matrices
4.3.7. Determine if there are any rare classes (< 30 field plots) and decide how to treat such rare classes.
4.3.7.1. Option 1: drop rare classes and re-run decision tree models.
4.3.7.2. Option 2: re-run decision tree models without the rare classes and then “burn” rare class
field plots onto the map.
4.3.7.3. Option 3: merge rare classes with floristically similar classes (solicit feedback from Veg- etation Working Group).
4.3.7.4. Option 4: retain the rare classes in the map.
4.3.8. Determine if other major mapping errors exist and correct by altering input parameters (if possible)
as well as field-referenced data.
4.3.9. Apply water, urban, and agriculture masks to life form-specific CT maps. 4.3.10. Merge the 3 life form-specific CT layers to form one CT layer.
4.4. Map life form-specific canopy height (CH)
4.4.1. Assign life form-specific CH classes to plots in modified MAT (4.1.3 and 4.1.4).
4.4.2. Extract digital values from the spatial layers, including life form-specific CTs (4.3.10), and use field plots classified to CH class values from 4.4.1 above.
4.4.3. Run decision tree models separately for the three life forms (tree, shrub, and herbaceous).
4.4.4. Generate life form-specific cross-validation error matrices for CH classes. 4.4.5. Generate life form-specific CH class maps using decision trees.
4.4.6. Check for errors in the three life form-specific CH maps, ensuring ranges of CH values are logical
for their corresponding CTs.
4.4.7. Mask each CH map with water, urban, and agriculture masks.
4.5. Map life form-specific canopy cover (CC)
4.5.1. Map tree CC
4.5.1.1. Create training set of forest CC using 1-m digital ortho-photography quadrangles or 1-m satellite imagery.
4.5.1.2. Establish the relationship between Landsat data and plot data using regression trees. 4.5.1.3. Apply the regression-tree relationship to generate a spatial per-pixel estimate of tree
canopy for all pixels.
4.5.1.4. Generate cross-validation error matrices, evaluate error and R2 values, and determine ef- fectiveness of the regression tree models.
4.5.1.5. Recode continuous tree CC data to CC classes defined by the Vegetation Working Group.
4.5.1.6. Apply land cover masks: water, urban, and agriculture. 4.5.2. Map shrub and herbaceous CC, option 1:
4.5.2.1. Extract digital values from the spatial layers using field plots that have shrub or herbaceous CC associated with them. Use the modified MAT (4.1.3 and 4.1.4).
4.5.2.2. Stratify digital values based upon dominant life form and run regression models.
4.5.2.3. Generate life form-specific error assessments based on cross-validation analysis.
Appendix 2-A — (Continued)
4.5.2.4. Determine effectiveness of the regression tree models based on error analysis and deter-
mine whether changes need to be made to both field data and independent spatial layers. 4.5.2.5. Generate life form-specific CC maps by applying the regression tree models.
4.5.2.6. Recode continuous variables to CC classes defined by the Vegetation Working Group.
4.5.2.7. Apply land cover masks: water, urban, and agriculture. 4.5.3. Map shrub and herbaceous CC, option 2:
4.5.3.1. Recode plot CC values in modified MAT (4.1.3 and 4.1.4) into CC classes defined by
Vegetation Working Group.
4.5.3.2. Extract digital values from the spatial layers (4.2.4) using binned shrub or herbaceous field
plots from step 4.5.3.1.
4.5.3.3. Stratify digital values based upon dominant life form (shrub and herbaceous vegetation) and run decision tree models.
4.5.3.4. Generate life form-specific error assessments based on cross-validation analysis.
4.5.3.5. Determine effectiveness of the decision tree models based on error analysis and determine
whether changes need to be made to both field data and independent spatial layers. 4.5.3.6. Generate life form-specific CC layers by applying the decision tree models.
4.5.3.7. Apply land cover masks: water, urban, and agriculture. 4.5.4. Map shrub and herbaceous CC, option 3:
4.5.4.1. Measure field spectral bands (corresponding to Landsat red and NIR bands) from multiple shrub and grass sites and derive field NDVI values.
4.5.4.2. Estimate percent shrub and herbaceous CC for sites where field spectral data has been
acquired (1-m2).
4.5.4.3. Determine relationship between field percent CC estimates and field-measured NDVI
values.
4.5.4.4. Estimate continuous shrub and grass CC through application of relationship described in
step 4.5.4.3 to Landsat NDVI to standardize Landsat CC estimates (stratified by life form
using NLCD 2000 data and/or LANDFIRE CT data).
4.5.4.5. Recode continuous shrub or herbaceous variables to CT classes defined by the Vegetation
Working Group.
4.5.4.6. Apply land cover masks: water, urban, and agriculture.
4.5.5. Refine and normalize CC estimates.
Appendix 2-A — (Continued)
4.5.5.1. Normalize individual tree, shrub, and herbaceous CC values such that tree, shrub, and herbaceous CC values combined do not exceed 100% per pixel.
4.5.5.2. Locate zones of low confidence using confidence layers and other sources of information. 4.5.5.3. Mask out zones of low confidence for shrub and grass CTs where forest is the dominant CT.
4.6. Generate merged CT and SS maps
4.6.1. Revisit, and revise if necessary, the merged CT map (4.3.10) by using forest, shrub, and herba- ceous percent CC as reference. Ensure that CTs match life form CC maps.
4.6.2. Produce a Federal Geographic Data Committee (FGDC) -compatible metadata file for the final
merged CT map (4.6.1) using a mapping zone metadata template for CT.
4.6.3. Generate a single CH layer using the CT data layer (4.6.1) for life form masking stratification. 4.6.4. Produce an FGDC-compatible metadata file for the final merged Canopy Height layer (4.6.3) using
a mapping zone metadata template for CH.
4.6.5. Generate a single CC layer using the CT layer (4.6.1) for life form masking stratification.